Hold a hotkey, speak, release, and the text lands where you were typing: a chat, an editor, a terminal, any app. Speech is recognized on your computer, on the GPU or a Mac's Neural Engine: no account, no subscription, and your voice never leaves the machine. Wavo lives in the menu bar or the tray.
Wavo is inspired by Handy and runs the same GGUF models, the ones handy-computer publishes. I built it for the technology: with agents around, there's no reason to put up with apps that aren't native and waste memory and compute.
- Handy is a webview app (Tauri) on its own engine, transcribe.cpp. Wavo draws a native UI on each OS with nimbo, light on memory, and runs its own engine,
wavo-cli, in pure Rust on the GPU. - On a Mac with macOS 15 or later, Wavo runs on the Neural Engine: 11 s of speech becomes text in 61 ms on an M2, 2.5× faster than on the GPU and 4.4× faster than transcribe.cpp, for about 6× less energy than the GPU (measured).
- It adds what I missed in Handy: an automatic microphone choice, and a history that can be exported with its recordings.
- Most alternatives are paid at the time of writing, and many are built on Whisper, whose last model came out in 2024. Wavo is free, works offline on macOS, Windows and Linux, and uses newer models, such as Parakeet Ultra.
- 🎙️ Hold to talk: hold the hotkey and speak, or tap it to start and stop; Esc cancels
- ⌨️ Pastes where you type: into any app, and the clipboard keeps what it held
- 🔒 Fully local: no account, no cloud; only the model is downloaded, once
- ⚡ Fast: 11 s of speech becomes text in 61 ms on an M2's Neural Engine (macOS 15 or later)
- 🌍 100 languages: Parakeet Ultra for 25 European ones, Whisper for the rest
- 💊 A quiet indicator: a small pill shows the level while you speak and never takes the focus
- 🎧 Picks the microphone: a USB one first, then the built-in one, a Bluetooth headset only when there is nothing else
- 📜 History: every dictation with its recording, searchable; play it back, transcribe it again, export it
- 📦 Shared models: kept in the standard Hugging Face cache, so other tools that use the same models share them
- 🖥️ Every desktop: macOS, Windows and Linux on X11 and Wayland, in about 10 MB
-
Download the file for your system from the latest release and run it.
-
Pick a model in Settings › Models: Parakeet Ultra, about 0.7 GB, downloaded once.
Which model? Parakeet Ultra is the default: one model for 25 languages with the lowest error rate of them. If you speak one or two languages, smaller single-language models can be faster and as accurate: I dictate with Parakeet v2 for English and GigaAM v3 for Russian, and switch between them in the menu.
-
Hold Option-Space (Ctrl+Alt+Space on Windows and Linux), speak and let go. A short tap starts a longer dictation that the next press ends; Esc cancels.
Coming from Handy? Quit Handy and run this, then bring in ~/Downloads/handy-export.zip with Advanced › Data › Import…: your history comes over with its recordings, and Handy itself is only read.
curl -fsSL https://raw.githubusercontent.com/vladkens/wavo-app/main/scripts/handy.py | python3 -The same Export… and Import… in Advanced › Data move your history to another computer.
Which file, and the first launch
ARM64 is Apple Silicon and other ARM chips; AMD64 is a usual Intel or AMD PC. <version> is the release's, such as v0.1.0.
- macOS 14.2 or later, ARM64:
Wavo-<version>-macos-aarch64.dmg; open it and drag Wavo into Applications. It is signed and notarized by Apple, so it opens like any other app. - Windows 10 or 11, AMD64:
Wavo-<version>-windows-x64-setup.exe; ARM64:Wavo-<version>-windows-arm64-setup.exe. It installs Wavo for you alone, no administrator needed, with a Start menu entry. It is not signed yet: SmartScreen says "Windows protected your PC", More info → Run anyway. Its icon is in the notification area, perhaps under the ^ arrow. - Linux AMD64:
Wavo-<version>-linux-x86_64.AppImage; ARM64:Wavo-<version>-linux-aarch64.AppImage. Make it executable (chmod +x) and run it. On Debian and Ubuntu the.debof the same name installs it instead:sudo apt install ./Wavo-<version>-linux-x86_64.deb.
The dev pre-release has the newest build of main, its files named Wavo-dev-….
Linux: Wayland, GNOME and the tray
On Wayland the desktop asks once to let Wavo paste (at the first paste, or earlier from the menu's "Allow Paste…") and once for the hotkey (from the menu's "Allow the Hotkey…"). The hotkey needs a desktop that offers global shortcuts: KDE Plasma 6, GNOME 48 and later, Hyprland. Elsewhere, as on GNOME before 48 (Ubuntu 24.04) or Sway, add a shortcut in the desktop's keyboard settings that runs /path/to/Wavo-<version>-linux-<arch>.AppImage toggle (with the .deb, wavo-app toggle): one press starts a dictation, the next stops it.
Wavo's icon goes in the tray, which GNOME needs the AppIndicator extension for (Ubuntu ships it); KDE, Xfce, Cinnamon and MATE have a tray of their own.
The GPU needs a Vulkan driver (mesa-vulkan-drivers on Debian and Ubuntu); without one Wavo runs on the CPU, much slower.
Wavo is built with AI agents following agents.md: I choose what to build, review the changes and test them on real machines. Before agents, I wrote software by hand for years (macmon, twscrape, ecloop). Releases are rare on purpose: I'd rather test and polish a set of changes than ship each one as it lands, and I try not to break your config or data; when something has to break, changelog.md says so.
Pull requests are welcome, including ones made with agents: point yours at agents.md.
Distributed under the MIT License.